UB55 multivariate analysis: hi-lo contrast, group-level
michael freund
2020-11-02
1 intro
Purpose:
- Preliminary look at hi/lo multivariate contrast (cross-validated euclidean) in each task.
- Preliminary qualitative comparison of effects of prewhitening and pattern standardization on multivariate contrast.
Notes on analyses:
Target TRs:
- Axcpt: 7, 8, 9
- Cuedts: 9, 10
- Stern: 11, 12
- Stroop: 2, 3, 4
Contrasts (cross-validated euclidean):
- Axcpt: \(\text{BY} - \text{BX}\)
- Cuedts: \((\text{InConSwitch} + \text{InConRepeat} - \text{ConSwitch} - \text{ConRepeat})/2\)
- Stern: \(\text{LL5RN} - \text{LL5NN}\)
- Stroop: \((\text{PC50InCon} + \text{biasInCon} - \text{PC50Con} - \text{biasCon})/2\)
Transformations:
- “whitened”: spatially pre-whitened cross-validated euclidean (‘mahalanobis’)
- using vertex-by-vertex inverse covariance matrix derived from residual timeseries
- “unwhitened”: not spatially pre-whitened
- all results considered here use unstandardized euclidean distance (i.e., patterns are not z-score normalized).
Plotting and statistical details:
- within ‘curves’ sections, ribbons indicate 95% CI of bootstrapped mean beta across subjects (one ribbon per run).
- all p-value corrections were performed whole-cortex, with FDR.
- for “stats and maps” sections, HLM was used to estimate t-statistics on the hi-lo contrast per ROI:
- random-intercept models
- level-I: (TR within target window)*(run)
- level-2: subject
- \(\beta_{ts} = b_0 + d_s + \epsilon_{trs}\)
- $_{ts} is HiLo contrast per subject, tr
- \(t, s\) are indices for TR, and subject
- \(d\) is is subject-specific intercept
- \(\epsilon\) is error
2 Network-level contrasts
Estimates are first averaged across parcels within network per subject. Group-level stats and plots are then displayed at the network-level.
2.1 beta estimates
2.1.1 t-stats
3 Parcel-level contrasts
3.1 counts of significant effects per TR
Using t-test of subject parcel means.
3.2 DMCC55B univariate ROIs
Curves for conjunction ROIs defined in univariate_group.rmd.
3.2.1 beta estimates
3.2.2 t-statistics
3.3 DMCC55B multivariate ROIs: Cross-task conjunction
Here conjunction analysis is performed by feeding TR-level contrast estimates to an HLM (see description at beginning of doc).
Number of parcels identified, and parcel names:
- No parcels ID’d (pFDR<0.05) in all four tasks in either whitened or non-whitened measures.
- Parcels ID’d in =>2 tasks with unwhitened:
## [1] "LH_Cont_Par_1" "LH_Cont_Par_2"
## [3] "LH_Cont_Par_3" "LH_Cont_Par_4"
## [5] "LH_Cont_Par_5" "LH_Cont_Par_6"
## [7] "LH_Cont_pCun_1" "LH_Cont_PFCl_1"
## [9] "LH_Cont_PFCl_2" "LH_Cont_PFCl_5"
## [11] "LH_Cont_PFCl_6" "LH_Cont_PFCl_7"
## [13] "LH_Cont_PFCl_8" "LH_Cont_PFCmp_1"
## [15] "LH_Cont_PFCv_1" "LH_Default_Par_5"
## [17] "LH_Default_Par_6" "LH_Default_PFC_1"
## [19] "LH_Default_PFC_10" "LH_Default_PFC_12"
## [21] "LH_Default_PFC_13" "LH_Default_PFC_17"
## [23] "LH_Default_PFC_20" "LH_Default_PFC_21"
## [25] "LH_Default_PFC_24" "LH_Default_PFC_5"
## [27] "LH_Default_PFC_7" "LH_DorsAttn_FEF_1"
## [29] "LH_DorsAttn_FEF_2" "LH_DorsAttn_FEF_4"
## [31] "LH_DorsAttn_Post_10" "LH_DorsAttn_Post_12"
## [33] "LH_DorsAttn_Post_4" "LH_DorsAttn_Post_6"
## [35] "LH_DorsAttn_Post_7" "LH_DorsAttn_Post_9"
## [37] "LH_DorsAttn_PrCv_1" "LH_SalVentAttn_FrOperIns_3"
## [39] "LH_SalVentAttn_FrOperIns_5" "LH_SalVentAttn_FrOperIns_7"
## [41] "LH_SalVentAttn_FrOperIns_9" "LH_SalVentAttn_Med_1"
## [43] "LH_SalVentAttn_Med_4" "LH_SalVentAttn_ParOper_2"
## [45] "LH_SalVentAttn_ParOper_4" "LH_SomMot_9"
## [47] "LH_Vis_22" "LH_Vis_28"
## [49] "LH_Vis_30" "RH_Cont_Par_2"
## [51] "RH_Cont_Par_3" "RH_Cont_Par_5"
## [53] "RH_Cont_Par_6" "RH_Cont_PFCl_10"
## [55] "RH_Cont_PFCl_11" "RH_Cont_PFCl_12"
## [57] "RH_Cont_PFCl_13" "RH_Cont_PFCl_14"
## [59] "RH_Cont_PFCl_4" "RH_Cont_PFCl_5"
## [61] "RH_Cont_PFCl_6" "RH_Cont_PFCl_7"
## [63] "RH_Cont_PFCl_9" "RH_Cont_PFCmp_2"
## [65] "RH_Cont_PFCv_1" "RH_Default_Par_4"
## [67] "RH_Default_PFCdPFCm_10" "RH_Default_PFCdPFCm_11"
## [69] "RH_Default_PFCdPFCm_12" "RH_Default_PFCdPFCm_13"
## [71] "RH_Default_PFCdPFCm_9" "RH_Default_PFCv_3"
## [73] "RH_Default_PFCv_4" "RH_DorsAttn_Post_8"
## [75] "RH_DorsAttn_PrCv_1" "RH_SalVentAttn_FrOperIns_5"
## [77] "RH_SalVentAttn_FrOperIns_8" "RH_SalVentAttn_Med_1"
## [79] "RH_SalVentAttn_Med_4" "RH_SalVentAttn_PFCl_1"
## [81] "RH_SalVentAttn_TempOccPar_3" "RH_SalVentAttn_TempOccPar_5"
## [83] "RH_SalVentAttn_TempOccPar_7" "RH_SomMot_7"
## [85] "RH_Vis_20"- Parcels ID’d in =>2 tasks with whitened:
## [1] "LH_Cont_Par_3" "LH_Cont_PFCl_6"
## [3] "LH_Cont_PFCl_8" "LH_Cont_PFCmp_1"
## [5] "LH_Default_Par_5" "LH_Default_Par_7"
## [7] "LH_Default_PFC_13" "LH_Default_PFC_24"
## [9] "LH_Default_PFC_7" "LH_DorsAttn_PrCv_1"
## [11] "LH_SalVentAttn_FrOperIns_5" "RH_Cont_PFCl_7"
## [13] "RH_Default_Par_1" "RH_Default_PFCv_4"
## [15] "RH_SalVentAttn_FrOperIns_5" "RH_SalVentAttn_TempOccPar_7"
## [17] "RH_SomMot_2"3.3.1 beta estimates
3.3.2 t-statistics
3.4 all parcels
3.5 maps (target TRs)
Unthresholded and thresholded t-statistics displayed on inflated brains. Threshold at fdr-corrected p-value (across all parcels) of 0.05. Note: colors are reverse-scaled, i.e., dark == strong effect, light == weak effect (so strong effects show up on white background).
3.5.1 Axcpt
3.5.1.1 whitened
3.5.1.2 unwhitened
3.5.2 Cuedts
3.5.2.1 whitened
3.5.2.2 unwhitened
3.5.3 Stern
3.5.3.1 whitened
3.5.3.2 unwhitened
3.5.4 Stroop
3.5.4.1 whitened
3.5.4.2 unwhitened
4 Task Profiles across parcels
This is done with DMCC55B univariate ROIs, for comparability with univariate analysis.
